Research map / Detection and segmentation / Open-set and few-shot detection
GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GTPBD, a global dataset of 47,537 high-resolution images with three-level labels (mask, boundary, parcel) for terraced agricultural parcels, and benchmarks segmentation, edge detection, parcel extraction, and UDA methods on it.
Problem
Existing agricultural parcel datasets focus on mid-resolution imagery or regular plain farmland and lack fine-grained annotations of complex terraced terrains, shared/non-shared boundaries, and multi-domain transfer settings.
Approach
A manually annotated, high-resolution (resampled to 0.5–0.7 m) global terraced parcel dataset spanning China and other countries, with pixel-level boundary, mask, and parcel labels, split into South/North/Global domains to support UDA benchmarking alongside pixel- and object-level metrics.
Key results
- NonLocal reaches the highest precision (75.06%) and OA (79.52%) among eight segmentation methods, while Mask2Former achieves the best recall (74.33%), IoU (57.16%) and F1 (72.74%).
- REAUNet-Sober is the best edge detector with ODS 65.06%, OIS 63.73%, AP 70.09%.
- HBGNet achieves the best parcel-extraction pixel metrics (IoU 62.44%, F1 76.88%), while Mask2Former gives the lowest object-level GOC (22.04%) and GTC (35.53%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Mask2Former instead of the other evaluated segmentation models | semantic segmentation on GTPBD · GTPBD | IoU | Helped |
| use REAUNet-Sober instead of the other evaluated edge detectors | edge detection on GTPBD · GTPBD | ODS | Helped |
| use HBGNet instead of the other evaluated parcel extraction models | parcel extraction on GTPBD · GTPBD parcel extraction | IoU | Helped |
| use Mask2Former instead of the other evaluated parcel extraction models | parcel extraction on GTPBD · GTPBD parcel extraction | GTC | Helped |
| use PiPa instead of the other evaluated UDA methods | UDA segmentation · six domain transfer directions, average | IoU | Helped |
| integrate REAUNet-Sober boundary results into UDA (ablation) | UDA methods on GTPBD · UDA Avg | IoU: 59.62 -> 59.13 | Hurt |
| integrate REAUNet-Sober boundary results into UDA (ablation) | UDA methods on GTPBD · UDA Avg | GTC: 36.91 -> 36.91 | Helped |
Surprising findings
- Source-only baselines are competitive with or outperform several UDA methods on some transfer directions (e.g., S→G and G→S precision), and UDA gains are inconsistent across metrics.
Limitations
- Terraced parcel annotations are manual and the dataset covers only 885 km² across 14 countries, limiting global representativeness.
- Benchmarks use default MMSegmentation-style training (20k iterations, SGD) without error bars or repeated runs.
Experimental setup
Datasets: GTPBD
Benchmarks: GTPBD
Metrics: Precision Recall IoU OA F1-score GOC GUC GTC ODS OIS
Baselines: U-Net DeepLabV3 PSPNet NonLocal OCRNet K-Net SegFormer Mask2Former MuGE PiDiNet UEAD REAUNet-Sober
Builds on: DAFormer HRDA PiPa REAUNet SEANet HBGNet
Closest papers by meaning
- Fields of The World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Segmentation (AAAI 2025 · similarity 0.87)
- Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.86)
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.85)
- HarvestNet: A Dataset for Detecting Smallholder Farming Activity Using Harvest Piles and Remote Sensing (AAAI 2024 · similarity 0.85)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.85)
- FUSU: A Multi-temporal-source Land Use Change Segmentation Dataset for Fine-grained Urban Semantic Understanding (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.